Andrew points out that 'x86/shadow: suppress "fast fault path"
optimization without reserved bits' assumes firm knowledge of the
physical machine's address width. When we run virtualized
ourselves, we can't reasonably assume that we do, the more that
the property may change as we may get migrated.

Since we want to further refine this logic anyway, I thought I'd
also include the other change that I've previously mentioned, at
least for consideration whether to take for 4.15. This reduces
the performance impact (albeit perhaps only for very large or
exotic guests) that the earlier patch has.

1: suppress "fast fault path" optimization when running virtualized
2: encode full GFN in magic MMIO entries

Jan